RESERVOIR PERFORMANCE PREDICTION USING TIME SERIES FORECASTING: AN AUTOMATED MACHINE LEARNING APPROACH

ABSTRACT

Reservoir performance prediction is an important part of oil and gas field development planning and reserve estimation because it depicts the reservoir's future behavior. Its success is dependent on reliable indicators of reservoir rock properties, fluid properties, rock-fluid properties, and flow performance. As a result, engineers must have a thorough understanding of reservoir characteristics and production operations optimization and the ability to develop a mathematical model that accurately depicts the physical processes taking place in the reservoir, allowing the outcome of any action to be predicted within permissible engineering tolerance levels. While there is no one-size-fits-all theory for production rate decline, there are broad laws and principles derived from historical data that can be used with sophisticated computer modeling to generate reasonably accurate production rate predictions. But then again, there is usually a lot of iteration involved in traditional machine learning approaches, as this usually would require trying out different machine learning models for accuracy and error margin. Eventually, this could prove really stressful and tasking. Hence, the need for this study of how reservoir performance prediction using machine learning time series analysis models can be automated. AutoML in Azure Machine Learning was used to carry out this prediction. It iterated through about 11 time-series models in the first round of predictions. In the second round of predictions, AutoML iterated over 35 models. Out of these models, VotingEnsemble was found to be the most accurate of them with an error score of 0.1488 in approximately two and a half minutes, in the first round of predictions and 0.1406 in the second round of prediction. Visualization was carried out using the pyplot library to see that the predicted values for monthly oil Ⅸ production were somewhat similar to the original values obtained from the dataset. This tells us the importance of AutoML in eliminating time constraints and also identifying deep-seated trends in the data that may not be explicitly visible to the eye during forecasting.